iraklixyz/georgian-wikipedia-cleaned

Cleaned Georgian Wikipedia Dataset (Ultra-Pure for LLM Pre-training)

A high-fidelity, ultra-pure, and thoroughly cleaned Georgian Wikipedia corpus derived from the latest XML dump. This dataset is specifically designed and filtered to contain 100% pure Georgian Mkhedruli text with zero foreign alphabetic pollution, making it ideal for vocabulary expansion, continued pre-training, or fine-tuning of Georgian Large Language Models (LLMs).

Dataset Summary

Wikipedia is a vital source of factual knowledge for training language models, but raw dumps are heavily cluttered with MediaWiki templates, list indices, stub-stamps, calendar pages, and external bibliography tables. Furthermore, standard dumps contain highly mixed multilingual content (inline English translations, Greek citations, or Cyrillic texts) which can degrade a Georgian model's tokenizer usage and prose learning weight.

This dataset applies a strict high-purity preprocessing pipeline to ensure absolute linguistic clean-up:

• Strict Georgian Ratio: Keeps only pages containing $\ge 95%$ Georgian Mkhedruli letters of all alphabetic characters.
• 100% Foreign Letter Stripping: Actively deletes all non-Georgian alphabetic letters (including Latin, Cyrillic, Greek, etc.) from the corpus text. Any resulting parenthetical remnants (e.g. empty inline brackets ()) are completely resolved.
• Prose-Only Filtering: Calendar stubs, numerical year pages, and list articles are completely excluded.
• Length Quality: Only pages with at least 300 characters of clean prose are retained.

The result is a pristine dataset containing 110,038 high-quality Georgian articles.

---

Dataset Structure

The dataset contains a single train split in a highly compressed Parquet format with the following schema:

Column	Type	Description
id	int64	Wikipedia Page ID
title	string	Title of the Wikipedia Article
text	string	Cleaned, ultra-pure Georgian prose text of the article
url	string	Wikipedia Article URL

---

Preprocessing & Filtering Pipeline

The raw XML dump was parsed and subjected to a multi-stage filtering pipeline to ensure absolute quality and purity:

1. Exclusion of Non-Prose Pages:
   • Calendar Days: Discarded standard calendar day stubs (e.g., 25 იანვარი) containing lists of dates/events rather than prose.
   • Year Articles: Discarded year pages (e.g., 1928), which are list-heavy timelines.
   • List Indices: Discarded list index pages (e.g. titles starting or ending with სია or სიები) to avoid training on fragmented nouns.
2. Strict Language Quality & Density Thresholds:
   • Language Ratio Filter: Enforced that at least 95% of all letters in each article must be Georgian Mkhedruli letters ([\u10D0-\u10FA]). This filters out foreign language pages, code snippets, or mathematical formula stubs. (15,199 pages skipped)
   • Stub Discarding: Enforced a strict minimum of 300 characters of clean prose. (51,950 raw pages and 150 post-strip pages skipped)
3. Prose Normalization & Absolute Purity Polish:
   • 100% Foreign Letter Stripping: Actively stripped all non-Georgian alphabetic letters (Latin, Cyrillic, Greek, etc.) from the text of all retained articles, leaving exclusively Georgian Mkhedruli characters, numbers, standard punctuation, and spaces.
   • Footer & Bibliography Truncation: Standard bibliography sections, footnotes, references, external links, and index footers (e.g., ლიტერატურა, რესურსები ინტერნეტში, სქოლიო, წყაროები, იხილეთ აგრეთვე) were located and stripped along with all subsequent text.
   • Punctuation & Paren Cleanup: Empty parentheses or relics left behind by template stripping (e.g., () or (; )) were automatically resolved.
   • Spacing Normalization: Whitespace and multiple tabs were collapsed to a single space, and consecutive newlines were restricted to a maximum of 2.

Pipeline Statistics

• Total Raw Inspected: 181,218
• Ultra-Pure Articles Retained: 110,038 (60.7% of raw dump)
• Dataset Size: 196 MB (196.03 MB, Snappy-compressed Parquet)
• Linguistic Purity: 100% pure Georgian Mkhedruli alphabetic characters (0% Latin/Cyrillic/Greek letters).

---

How to Use

You can load this dataset directly in Python using the Hugging Face datasets library:

from datasets import load_dataset

# Load the dataset
dataset = load_dataset("iraklixyz/georgian-wikipedia-cleaned")

# Inspect a sample
sample = dataset["train"][0]
print(f"Title: {sample['title']}")
print(f"URL: {sample['url']}")
print(f"Text Preview:\n{sample['text'][:500]}")

---

Creator Attribution & Collaborations

This dataset was created and preprocessed by Irakli Maisuradze (@iraklixyz).

If you have suggestions, find issues, or want to contribute to the scraping and cleaning tools, feel free to open an issue or pull request on the GitHub repository!

---

Citation

If you use this dataset in your research, language models, or academic publications, please cite it using the following BibTeX format:

@misc{georgian_wikipedia_cleaned,
  author    = {Irakli Maisuradze (iraklixyz)},
  title     = {Cleaned Georgian Wikipedia Dataset for LLM Pre-training},
  year      = {2026},
  publisher = {Hugging Face},
  journal   = {Hugging Face Datasets},
  howpublished = {\url{https://huggingface.co/datasets/iraklixyz/georgian-wikipedia-cleaned}}
}

---

License

This dataset is distributed under the same license terms as the source material: Creative Commons Attribution-ShareAlike 4.0 International (CC BY-SA 4.0).